Local density based similarity criterion for clustering of remote-sensing images

Unsupervised clustering is a powerful method for land cover identification using remote-sensing images. Due to increasing spatial resolution and improved satellite capabilities, these images have had large sizes, which in turn makes pixel-based clustering often infeasible and necessitates prototype-based clustering. The use of prototypes comes with advantages such as robustness to noise and outliers, but more importantly, new types of information for similarity definition in addition to distance-based approaches. A recently proposed local density based similarity (CONN) is shown powerful for hierarchical and spectral clustering. This study shows its success in clustering of remote-sensing images for agricultural monitoring.